Gorse 0.5.x Config Migration from 0.4

SkillMedia

Fix Gorse 0.5.x recommendation engine silently disabled or degraded due to config incompatibility with 0.4. Use when: (1) Gorse 0.5 recommendations are poor/empty despite having data, (2) collaborative filtering appears disabled even though [recommend.collaborative] section exists, (3) migrating Gorse config from 0.4 to 0.5, (4) old config keys like [recommend.popular], [recommend.neighbors], [recommend.online], [recommend.offline] are being used with Gorse 0.5. Critical: type="mf" must be explicitly set in [recommend.collaborative] or CF is silently disabled (default "none").

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Gorse 0.5.x Config Migration from 0.4 skill

What this skill tells your AI

The instructions your AI receives, as published by divinevideo/divine-mobile in .agents/skills/gorse-05-config-migration/SKILL.md and read by ahel’s review.

Problem

Gorse 0.5 introduced breaking config changes from 0.4. Old config sections are silently ignored (no error, no warning), making it appear that features like collaborative filtering, trending, and item neighbors are configured when they're actually disabled.

The most critical issue: [recommend.collaborative] defaults to type = "none" in 0.5, meaning collaborative filtering is silently disabled unless you explicitly set type = "mf".

Context / Trigger Conditions

  • Gorse 0.5.x is deployed but recommendations seem random or low quality
  • Collaborative filtering doesn't seem to work despite config having [recommend.collaborative]
  • Config has old 0.4 sections: [recommend.popular], [recommend.neighbors], [recommend.online], [recommend.offline]
  • Upgrading from Gorse 0.4 to 0.5
  • Gorse master starts successfully but old config features don't take effect

Solution

Removed 0.4 Sections (silently ignored in 0.5)

0.4 Section0.5 Replacement
[recommend.popular][[recommend.non-personalized]] (TOML array of tables)
[recommend.neighbors][[recommend.item-to-item]] + [[recommend.user-to-user]]
[recommend.online] (explore/exploit)[recommend.ranker] (FM or LLM-based)
[recommend.offline][recommend.ranker] cache_expire + [recommend.fallback]

Removed 0.4 Keys (do not exist in 0.5)

  • enable_item_neighbor_index
  • enable_user_neighbor_index
  • item_neighbor_type
  • neighbor_type
  • explore_recommend
  • popular_window

Critical Fix: Enable Collaborative Filtering

# 0.4 style (BROKEN in 0.5 — CF silently disabled):
[recommend.collaborative]
fit_period = "60m"
fit_epoch = 100

# 0.5 style (WORKING — must add type = "mf"):
[recommend.collaborative]
type = "mf"
fit_period = "60m"
fit_epoch = 100

Complete 0.5 Config Template

[recommend]
cache_size = 100
cache_expire = "72h"
context_size = 100

[recommend.data_source]
positive_feedback_types = ["reaction", "comment", "repost"]
read_feedback_types = ["view"]
positive_feedback_ttl = 0
item_ttl = 0

# Collaborative filtering — MUST set type = "mf" to enable
[recommend.collaborative]
type = "mf"
fit_period = "60m"
fit_epoch = 100
optimize_period = "360m"
optimize_trials = 10

[recommend.collaborative.early_stopping]
patience = 10

# Trending/popular — TOML array syntax [[...]] for multiple leaderboards
[[recommend.non-personalized]]
name = "trending_weekly"
score = "count(feedback, .FeedbackType == 'reaction') + count(feedback, .FeedbackType == 'comment') * 2"
filter = "(now() - item.Timestamp).Hours() < 168"

# Content-based item similarity (requires items to have Labels)
[[recommend.item-to-item]]
name = "similar_content"
type = "tags"         # matches on item Labels

# Collaborative item similarity
[[recommend.item-to-item]]
name = "also_liked"
type = "users"        # users who liked X also liked Y

# User similarity
[[recommend.user-to-user]]
name = "similar_users"
type = "items"        # users who share liked items

# FM ranker blends all candidate sources
[recommend.ranker]
type = "fm"           # or "llm" for LLM-based reranking
recommenders = ["latest", "collaborative", "non-personalized/trending_weekly", "item-to-item/similar_content", "item-to-item/also_liked", "user-to-user/similar_users"]
fit_period = "60m"
fit_epoch = 100

[recommend.ranker.early_stopping]
patience = 10

[recommend.fallback]
recommenders = ["latest"]

[recommend.replacement]
enable_replacement = true
positive_replacement_decay = 0.8
read_replacement_decay = 0.6

Key 0.5 Config Differences

  1. [[recommend.non-personalized]] uses Expr language for score/filter functions

    • Available: count(feedback, .FeedbackType == 'X'), item.Timestamp, now()
    • Filter: (now() - item.Timestamp).Hours() < N
  2. [[recommend.item-to-item]] types: "tags" (label similarity), "users" (collaborative), "embedding"

  3. [[recommend.user-to-user]] types: "items" (shared items), "tags", "embedding"

  4. [recommend.ranker] types: "none", "fm" (factorization machines), "llm"

    • recommenders list uses "category/name" format to reference specific recommenders
  5. Custom feedback types: Any string is valid in positive_feedback_types (e.g., "extended_view", "bookmark")

Verification

After updating the config:

  1. Restart Gorse master and check logs for config parse errors
  2. Visit Gorse dashboard (port 8088) — new recommenders should appear
  3. Check /api/dashboard/config endpoint to verify config was accepted
  4. After fit_period (default 60m), verify model training runs in worker logs

Notes

  • Gorse 0.5 does NOT warn about unrecognized config keys — they're silently ignored
  • The [[double.bracket]] syntax is TOML array of tables — allows multiple entries
  • The [single.bracket] syntax is a regular table — only one allowed
  • [recommend.ranker].recommenders must reference recommenders by their "category/name" path
  • Custom feedback types are supported — Gorse doesn't validate feedback type names

References

Signals

GitHub stars
265
Forks
55
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
gorse-05-config-migration
Source
github.com/divinevideo/divine-mobile